Meta Open-Weights an 816B Model, With a Revenue Line in the Licence

Meta published the weights for Muse Spark on 15 October, seven months after the model debuted as a closed product in April 2026. The download is real and the model is large: 816 billion total parameters with 42 billion active per token, a mixture-of-experts design that keeps the compute per token far below what the parameter count suggests.

The number that will decide who can build on it is not in the architecture. It is 20 million dollars, the gross-revenue threshold at which the new Muse Model License stops being free and starts requiring a separate agreement. The terms are set out in the announcement post, and the weights themselves are on Hugging Face.

What the licence permits, and where it stops

SiliconANGLE's read of the release is that there is no fee to adapt or embed the model, and that the limits arrive when you host it for other people. A provider offering Muse Spark Open as a service needs a separate agreement once it crosses 20 million dollars in gross revenue over twelve months. Products built on the model must carry a "Built with Muse Spark Open" label, and the licence prohibits military and critical-infrastructure use.

That structure is the point of the release as much as the weights are. Meta gets the distribution and the mindshare that come with an open model, while the business of selling inference on it stays behind a negotiation with Meta. This is a custom licence, not an OSI-approved open-source licence, and calling the model "open source" without that qualifier overstates what was published. Open weights are the accurate description.

The people most affected are the ones who would have been the strongest advocates. An inference provider with a healthy margin has to model a licensing cost into its pricing above the threshold, and the threshold is measured on gross revenue, not profit. A startup that hopes to resell the model in a managed product has one more thing to track. For an individual developer or a company running the model internally, none of these clauses apply.

The comparison the release invites is with Chinese open-weight models rather than with frontier APIs. Meta's own framing is about what a US lab can publish, and the Terminal-Bench row shows that on coding the answer is currently a Chinese one. A licence that restricts hosting is also a licence that complicates the audit story that attracts some teams to open weights in the first place: you can inspect and modify the model, and you can still be asked to remove it from a product if the terms are read differently than you read them.

The vendor benchmarks are strong in places and not in others

The headline claim, that this is the most capable open-weight model to come out of a US lab so far, is Meta's own. Its published table supports the first half of that sentence and is awkward about the second. Against the open models Meta chose to list, Muse Spark Open scores 61.8 on JobBench, where Inkling scores 31.1 and Nvidia's Nemotron 3 Ultra scores 27.2. On the MCP-Atlas tool-use evaluation it reports 89.2, described as the best of the models on the list, measured over more than 220 tools across 36 live MCP servers.

Coding benchmarks tell a different story. On Terminal-Bench 2.1 the model reports 76.4, below Kimi K3 at 85 and below GLM 5.3 and Qwen3.8 as well. A team choosing an open model for agentic coding should read that row carefully, because it is the row the licence does not flatter. All of these numbers are vendor-reported, and none of them were independently reproduced.

A model this size also changes the shape of an evaluation. Fine-tuning 816 billion parameters is not a workstation task, and even serving a quantised build for internal traffic needs memory that most teams do not have spare. For many organisations the honest reading of this release is that it is an option for their inference provider rather than for them, and the provider is the party that has to answer the licence question above the revenue threshold. Self-hosting becomes a procurement project rather than a download, which is a different kind of openness from the one the phrase suggests.

Running 816B is a hardware decision, not a download

Meta demonstrated the model on two Mac Studio M5 Ultra machines with roughly a terabyte of combined memory. That is a useful calibration for anyone reading "open weights" as "runs on my laptop". At this size the practical routes are a large-memory workstation, a rented multi-GPU node, or one of the seven inference providers announced at launch, a list that includes Groq and Fireworks AI. The weights are on Hugging Face and in the AWS, Google and Microsoft catalogues.

The context window runs up to one million tokens and the model accepts text, images, audio and video prompts. Long context changes the memory arithmetic again, because the KV cache grows with the conversation rather than sitting at a fixed cost. A million-token window is a capability on paper and a capacity-planning problem in practice, particularly for a self-hosted deployment where nobody else is absorbing the spike.

What to do with the release

Read the licence before you benchmark. The clauses that matter are the revenue threshold, the labelling requirement and the prohibited uses, and you need an answer on all three before a product ships rather than after it has customers.

Check the weights against your actual workload, not Meta's table. The gap between the JobBench lead and the Terminal-Bench deficit is wide enough that a coding agent and a tool-calling agent should reach different conclusions from the same release. Pull the model from Hugging Face, run it on your own traffic sample, and compare it against the Chinese open models it does not beat on coding.

If you do host it for others, diarise the revenue line. Twenty million dollars over twelve months is a threshold a growing inference business can cross without a deliberate decision, and the agreement it triggers is not one to negotiate after the invoice.